Computer Science (arXiv)

The integration of generative AI into web search delivers synthesized answers to user queries, changing how people navigate and assess information, while raising concerns about the downstream impacts on publishers who supply the underlying content. We conduct a preregistered field experiment (N=1,100) on Google Search, the dominant online search platform, to estimate the causal effects of AI Overviews and AI Mode on user behavior, perceptions, and publisher traffic. We show that removing AI Overviews and AI Mode increases click-through rates to publishers, while an AI Mode-only experience reduces click-through rates and erodes user experience and trust in information found on Google. These findings show that integrating generative AI into web search reshapes online attention, with economic consequences for the online publishers that sustain both search platforms and the overall information ecosystem.

Tool-using large language-model agents can complete a task while exercising authority that the user did not grant or the task does not need, causing excess-authority errors. Traditional permission gating systems alone for validating agent environments are insufficient. We study whether post-training can teach a 4B-parameter model to choose task-conditioned authority in executable terminal and Model Context Protocol (MCP) environments to complement those measures. We propose a framework where each action is audited before execution and again from observed effects along six dimensions of risk. This auditing is conducted using deterministic verifiers that score completion, evidence, exact state, prohibited attempts, and safe success. In conjunction with predefined task-specific sufficient-authority envelopes, we determine task-specific excess privilege values for trajectories, which are then optimized for in post-training. We find that after training using this framework on Qwen3.5-4B over 1,500 tasks, the selected seed reaches 98.48% safe success across 2,896 evaluation episodes spanning all 500 held-out tasks, compared with 64.36% for the base policy, and reduces excess-authority error events from 4.56% to 0.79%. Furthermore, external tests show capability retention and prompt-directed improvement. A 400 task continuation study also found evidence of generalization, reducing excess-authority events by 6.99 percentage points while maintaining previous capabilities. We conclude learned restraint through least-privilege aware post-training is therefore useful as an additional control layer for tool-using agents in executable terminal and MCP environments, but it does not replace permission gates and sandboxing.

Growing network speeds, with 100GbE line rates becoming common in modern enterprise networks, pose challenges to operators and security applications, as they struggle to scale their operational efficiency accordingly, without relying on costly hardware, excessive sampling, or complex distributed deployments. Unintentional loss due to stochastic packet sampling often produces low-quality traffic, further risking missed detection of critical security incidents, particularly those hidden in typically low-rate traffic, such as APT/malware command-and-control communications. In this paper, we introduce XNET, a system that monitors traffic at line rate using commodity hardware and applies dynamic sampling to amplify the visibility of high security value traffic. XNET leverages Linux's XDP technology to process packets efficiently, classify them based on their security value, and sample them as per configured policies. The outcome is a reduced packet stream in which the security-relevant portion of the traffic is amplified at the expense of less interesting traffic segments. XNET is a highly flexible, scalable and dynamic system that can be adapted based on a network's needs. We deployed XNET in a large real-world network using only commodity hardware, where our results show that XNET can achieve up to 84% traffic reduction with no packet loss while increasing the visibility of otherwise negligible traffic fivefold. With controlled stress tests, we further demonstrate XNET's scalability up to 100Gbps. Additionally, we show that XNET sampling led to a detection rate of 99.6% in an IDS application.

Coupled-cluster theory defines the accuracy standard for molecular electronic-structure properties but scales too steeply for routine application, whereas density-functional theory is affordable yet systematically biased. We resolve this trade-off with a single equivariant network, MEHnet-MG, that predicts an effective one-electron Hamiltonian from one inexpensive B3LYP/def2-SVP calculation and derives a broad suite of properties from it (energy, optical gap, dipole, quadrupole, polarizability, Mulliken atomic charges, and Mayer bond orders) at coupled-cluster accuracy across nine main-group elements, including the under-served phosphorus, sulfur, and chlorine chemistries. The model is trained on a new in-house dataset of multi-property labels computed at the CCSD(T) level for all nine elements. On a held-out test set, it reduces the error of every property by a factor of 3.8 to 230 relative to semi-local, hybrid, and double-hybrid DFT (referenced to composite CCSD(T)/cc-pVTZ; Methods), while adding only ~25 ms wall time per molecule, delivering coupled-cluster-quality predictions at the cost of a single DFT calculation. Critically, deriving every property from a predicted Hamiltonian rather than pooling per-atom features builds the correct size-scaling into the model architecture: on pi-conjugated oligothiophenes it matches finite-field CCSD polarizability and the EOM-CCSD optical gap to ~2% at the largest sizes where those references remain affordable (44 and 37 atoms, where a single CCSD field point already costs ~500x the model's entire inference) and extrapolates the corrected trends to 58-atom chains, a regime where pooling-based architectures fail by construction. Accurate extrapolation is therefore set by the model's inductive bias rather than by the training data.

Persistent acoustic monitoring can detect machine faults without physical contact, but always-on inference is constrained by power, latency, and deployment complexity. We demonstrate autoencoder-based acoustic anomaly detection on an Intel Loihi 2 neuromorphic processor under clean and noisy conditions. Log-mel features are computed off chip; normalization, autoencoder inference, L1 reconstruction scoring, and thresholding run on chip. In a clean, microphone-position-invariant ToyADMOS ToyCar benchmark, the on-chip model achieves 0.9959 AUC and 0.9785 standardized pAUC at maximum false-positive rate 0.1. In the DCASE 2026 Task 2 ToyCar noisy benchmark, the model achieves source AUC 0.7990, target AUC 0.6466, and pAUC 0.6426, exceeding reported baseline metrics. Power profiling on a 16-chip Loihi 2 VPX system shows real-time throughput with 0.0406$\unicode{x2013}$0.0426 mJ dynamic energy per sample, two orders of magnitude lower than both a CPU and GPU. These results support neuromorphic acoustic anomaly detection as a practical candidate for low-power, persistent machine monitoring.

Vision-language models (VLMs) have demonstrated remarkable zero-shot capabilities yet remain sensitive to real-world distribution shifts during inference. Although significant efforts are devoted to adapting VLMs at test time, they rely heavily on noisy pseudo-labels predicted directly from raw embedding similarities during inference, which are unreliable under distribution shift and mislead the adaptation. To avoid noise amplification, existing works craft coarse-grained surrogate objectives during adaptation, which fail to explicitly model sample-level relationships across different modalities, creating objective mismatch with inference, thus leading to marginal performance improvement. In this work, we aim to bridge the detached objectives of inference and adaptation for VLMs, and propose a principled VLM TTA method called \algname. For VLM inference, we formulate the zero-shot image classification task as a cross-modal alignment problem encoded via a Wasserstein OT formulation, providing robust pseudo-labels at the sample-level to effectively adapt VLMs. For VLM adaptation, we adopt a soft-label InfoNCE loss to adapt VLMs based on the OT-induced pseudo-labels, leveraging fine-grained supervisions to explicitly model relationships of individual image-text pairs via contrastive learning, which empowers accurate inference at the same granularity. Moreover, we theoretically reveal that the InfoNCE loss can be neatly reformulated as a Wasserstein OT formulation, thereby unifying the objectives of the inference and adaptation of VLMs to achieve their mutual benefits. Extensive experiments demonstrate the effectiveness and efficiency of our methods, outperforming the best-performing methods by up to 7% with state-of-the-art efficiency.

When evaluating language models on human exams, benchmarks typically score each response as right or wrong and report the overall accuracy. This approach assumes that partial knowledge is worth proportional credit, an assumption that fails when an examination uses a non-additive grading scheme. The 2025 reform of Vietnam's National High School Graduation Examination demonstrates the cost of this substitution. In Part II of the exam, candidates evaluate four true/false statements per question. The grading is convex: the number of correct statements earns 0, 0.10, 0.25, 0.50, or 1.00 points. Identifying three statements correctly pays 0.50 points, not the 0.75 points that standard accuracy metrics would award. Because Part II accounts for 4.00 of the exam's 10.00 points, reporting accuracy inflates the score by rewarding partial knowledge that the state explicitly penalizes. We introduce THPT-Ladder, a benchmark of 632 items from 21 official exams across 11 subjects, graded exactly as the ministry grades its students. The ministry publishes the marks of over a million candidates, allowing us to place models directly into the human cohort. Across eight models, the official rubric pays 0.020 to 0.159 points less per Part II question than proportional credit. This shortfall changes a model's apparent competence. For Qwen3.5-27B on the 2025 History exam, a 0.042-point shortfall drops its standing from the 90th to the 77th percentile among 481,293 candidates. A model's accuracy does not predict this penalty. At Claude Sonnet 5's accuracy level, different distributions of errors yield scores varying from 0.869 to 0.932 points per question. Official marks depend on how correct statements are grouped, meaning standard benchmarks report a competence the institution would not certify.

Whether input-dependent ("dynamic") combination of a regression model pool beats the best static blend depends on the shift and is rarely known before deployment. Can a small labeled target-domain probe tell us when reallocating trust across regions of the input space will pay off? We answer this with $\widehat{D}_{\mathrm{CF5}}$, which estimates from the probe the cross-fitted gain of the regionwise convex combination over the best static convex blend: the realizable value of deciding, region by region, whom to trust. Across a frozen suite of 12 dataset-shift pairs (spatial, temporal, domain, feature-cluster), $\widehat{D}_{\mathrm{CF5}}$ predicts realized regionwise test gains with dataset-level Spearman $+0.98$ (95% CI $[+0.83, +1.00]$; $p=5\times10^{-5}$), including two cases overturning preregistered expectations. The relationship holds in a 16-pair sensitivity analysis (Spearman $+0.83$), whereas alternative probe diagnostics reach at most $+0.66$. This contrast isolates regional trust reallocation: correlation is $+0.98$ for regionwise-convex gain, but $+0.01$ for smooth covariate-dependent stacking after affine correction. A controlled generator shows dynamic gains arise from the interaction of shift heterogeneity and local competence, increase with shift severity, and become realizable between 128 and 256 probe labels in the tested grid. The Probe-Validated Ensemble Selector chooses among a static affine stacker and dynamic realizers, deploying a candidate only when a held-out lower confidence bound clears the static-convex floor. In a preregistered prospective batch, it matched or improved the floor in all 12 runs; two deployments reduced test risk by 11% and 16%, while the gate rejected a candidate whose un-gated deployment incurred $>30\times$ the static loss. We release OpenRegShift, a reproducible evaluation harness for regression ensembles under distribution shift.

There is no shared framework for selecting among document compilation engines (pdfLaTeX, XeLaTeX, LuaLaTeX, Tectonic, Typst, and pandoc PDF backends). We present Engine-Transfer-Bench (ETB): 1,784 open documents, four tasks covering reliability, latency, text consistency, and failures, a pinned harness, and host-tagged multi-OS results. On GitHub Actions (N=4,211 compiles per host across macOS, Ubuntu, and Windows), Tectonic success is stable within 0.9 percentage points (96.3-97.2%), whereas classic TeX Live-style engines vary by 12-20 percentage points according to distribution policy (Ubuntu apt, MiKTeX auto-install, or macOS BasicTeX). On 702 portable LaTeX documents, the tested engines succeed at 100%, making latency the primary selection factor; failures concentrate in 107 engine-specific templates. Within ETB, failures are architectural, involving fonts, layout, and assets, rather than missing packages on a provisioned host. A 50-pair validation of the S_pdf text-consistency metric achieves 94% precision for real content divergence. We release ETB, ETB-Porta, a recommender and portability gate, and a public cross-OS harness as shared infrastructure.

Concolic execution is a variant of symbolic execution that runs a program simultaneously with concrete and symbolic inputs. It records the symbolic constraints encountered along a concrete execution path, then solves those constraints to generate inputs that explore new paths. Existing concolic engines generally follow one of two implementation strategies: Interpreter-based systems are comparatively simple to build but incur substantial interpretation overhead, while instrumentation-based systems avoid this overhead but typically re-execute the program from the beginning for each new input. In this paper, we develop a new approach that achieves the best of both worlds. Starting from the concrete semantics of the target language, we first develop a definitional concolic interpreter and stage it to compile away interpretation overhead while retaining the simplicity of an interpretation-based implementation. By expressing the staged interpreter in continuation-passing style, we can capture execution snapshots at branch points and resume from them when exploring alternative paths, avoiding repeated execution from the program entry. Because snapshot-reuse can itself incur overhead, we further develop a heuristic that favors snapshot-reuse only when it is expected to be beneficial. We instantiate this approach for WebAssembly and implement it in a new concolic-execution compiler GenWasym. Across 184 benchmarks, GenWasym with staging alone achieves a $29.4\times$ average speedup over the interpreter-based WASP; heuristic snapshot-reuse further increases the speedup to $44.9\times$.

Machine-verifiable workflows produce governance records linking a task contract, model attempt, verifier decision, accepted output, and target origin. We test whether these records can supervise bounded models, consolidating occasional or expensive capability into reliable one-shot execution. On fresh, structure-disjoint PlanBench replanning cases, Qwen3-14B thinking generated 24 plans admitted by the independently authored VAL verifier. Those plans trained the same checkpoint for non-thinking execution, without oracle targets or a stronger teacher. On 80 unopened cases, VAL-accepted plans increased from 1 to 57, with 56 paired gains and zero regressions; thinking reached 30. The adapter was schema-valid on all cases and used approximately 1/56 of thinking's mean latency. The separate paired interface-cure gate did not pass. A matched ablation fixed the source cases, 52-candidate pool, 24-target count, model, recipe, and seed while changing target selection. On 160 new cases, base, schema-selected, model-self-selected, and VAL-selected execution reached 1, 55, 69, and 102 accepted plans. VAL exceeded self-selection by paired net +33 (p=0.0000019647), with gains in both difficulty strata. Independent semantic selection is therefore load-bearing relative to matched alternatives within this band. A complementary Phi stronger-teacher arm raised base Phi-4 from 2 to 51 accepted plans and from 35 to 80 schema-valid outputs. Earlier synthetic experiments establish teachability, cumulative learning, construction robustness, and stopping boundaries. The results support verifier-selected supervision for bounded, machine-checkable capabilities, not arbitrary planning, enterprise validity, or unrestricted self-improvement.

Personalized promotional assets, namely artwork images and video preview clips, are critical to content discovery on Netflix. Traditional models for asset selection rely on ID-based interaction history, leaving them blind to asset content and unable to serve newly launched titles and assets. We describe how multimodal embeddings reshaped production systems at Netflix and report transferable lessons for practitioners adopting foundation-model embeddings into recommender systems. First, pretrained image embeddings unlock cross-title, cross-canvas knowledge transfer. Augmenting a two-tower model with CLIP image embeddings lets a single model serve all five Netflix artwork canvas types, replacing five separately trained per-canvas models and substantially improving cold-start performance. A lightweight extension reuses CLIP's joint text-image space to make artwork personalization query-aware in search. Second, multimodality decisively beats any single modality for video preview personalization. We describe MediaFM, our in-house tri-modal foundation model trained on a large-scale corpus of shots from the Netflix show catalog, fusing visual (SeqCLIP), audio (wav2vec 2.0), and timed-text signals; adopted for video preview personalization, it outperforms strong visual-only baselines both offline and in online A/B tests. Third, a simple offline proxy task whose performance correlates with online outcomes can accelerate the experimentation and productization cycle. Predicting the popularity-based winner from embeddings alone ranks embedding models and versions, pruning the choice space before any end-to-end integration or A/B test; it now gates every new MediaFM checkpoint. We also share the production engineering decisions (shared embedding infrastructure, low-latency serving, cheap screening) that made these deployments viable, along with the design tradeoffs and failure modes we encountered.

Neural networks trained on nonstationary tasks frequently lose the ability to fit new targets, a phenomenon referred to as loss of plasticity. We identify a novel source of plasticity loss due to the growing anisotropy of weight matrices' singular values during training, and analyze this phenomenon both empirically and theoretically. To mitigate this issue, we introduce SingularClip, a procedure that periodically clips the singular values of all weight matrices. We show that SingularClip performs strongly against baselines across a range of tasks in both continual supervised learning and deep reinforcement learning.

Affordance prediction is the identification of potential actions an agent can perform on a target object from multimodal inputs. Affordance prediction methods are difficult to evaluate and compare due to heterogeneous problem formulations, inconsistent dataset annotations, incomplete reporting of experimental protocols, and limited information about deployment conditions. These limitations challenge fair benchmarking and performance comparison. To promote transparency, we propose the Affordance Sheet, a documentation detailing task formulation with its input modalities, model architectures and training information, datasets, and experimental protocols. Affordance Sheets enable reproducible benchmarking and reliable evaluation of affordance models for real-world scenarios, including generalisation to novel conditions and human safety.

Structural Magnetic Resonance Imaging (MRI) is widely used in neuroimaging research and clinical practice, but structural MRI volumes may retain facial and cranial anatomical information that raises privacy concerns. Existing deep learning-based brain extraction methods generally produce a single fixed output, limiting flexibility when different applications require different balances between privacy and anatomical preservation. This paper presents a configurable privacy-preserving MRI processing workflow that extends deep learning-based brain extraction through adaptive anatomical preservation, interactive preservation selection, and integrated quality control. The workflow employs SynthStrip for automated brain extraction, followed by morphological mask expansion to generate configurable shell-based preservation levels. An Interactive Preservation Framework enables users to compare preservation configurations and select an appropriate output, while an integrated Quality Control Framework provides multi-plane visualisation and brain-mask overlay verification. The workflow was implemented in Python using open-source neuroimaging libraries within the Renku reproducible research environment and evaluated using structural T1-weighted MRI data from the publicly available IXI dataset. Experimental results demonstrate anatomically plausible brain extraction and configurable preservation outputs, supported by systematic visual verification. The principal contribution is a modular and reproducible MRI preprocessing framework that enhances deep learning-based brain extraction with configurable anatomical preservation, interactive user-guided processing, and integrated quality control. The workflow provides a practical foundation for privacy-oriented neuroimaging research and collaborative medical image analysis.

Autonomous scientific agents now increasingly propose ideas, write code, run experiments, analyze results, and even draft papers. Observe and audit those agents are necessary but logging every model call is not enough, scientists also need to inspect the artifacts and claims that the systems produced and their relations. This is driven by the fact that failures in scientific agent systems are often distributed across several objects. A manuscript claim may cite the wrong evidence, a search process may select a degenerate candidate, a laboratory novelty claim may depend on an unstated rule, or a multi-agent plan may change without a visible trigger. Existing tracing, experiment tracking, and archival provenance tools are valuable, but their native objects do not make these scientific audit relations first-class. We argue that autonomous scientific systems should emit portable, claim-aware artifact lineage as a minimum audit layer. We propose a compact observability profile organized around individuals, operators, fitness records, lineage, archives, runs, streams, and steering commands. In this profile, scientific claims are ordinary individuals with explicit evidence bindings and verification records. The profile is intended as a semantic layer that complements current telemetry and provenance standards. Execution details can remain in OpenTelemetry. Final packages can export to PROV-O or RO-Crate standards.

Federated multimodal models often assume every site has every modality, although hospitals differ in access to EHRs, chest radiographs, and ECGs. We study this setting on a MIMIC-derived respiratory deterioration task with simulated FL clients and introduce FedCoRe (Federated Cross-Modal Representation Completion). FedCoRe learns representation- or logit-space corrections rather than generating synthetic ECGs or CXR images. When a client observes a modality that may be missing at deployment, it evaluates the same example with and without that modality to obtain paired supervision. Only clients with such pairs update the completion module, and validation may retain the unchanged prediction. We freeze the trained multimodal predictor during evaluation so that measured differences come only from completion. Hiding ECG reduced AUROC by about 0.085; paired-example FedAvg restored 0.0415 AUROC, or 49.0% of the lost performance. We therefore report two distinct effects: paired-example FedAvg partially recovers the missing-ECG gap, while validation-selected completion is a task-specific classifier-logit correction rather than literal ECG recovery. For CXR, effect-aware completion recovers 52.8% of the loss in a controlled test where CXR is hidden. Paired-example FedAvg transfers part of this effect, but validation keeps the no-completion baseline for deployment cases whose inputs lack CXR. Thus, FedCoRe should be read as a validation-gated completion/correction framework: it can recover missing-modality signal in supported settings, but it should be deployed only when paired examples and validation evidence support that modality.

We propose UHDformer++, a general Transformer-based framework to solve numerous Ultra-High-Definition (UHD) image restoration tasks. UHDformer++ operates across $4$ coordinated learning spaces: 1) a high-resolution space (HR) for multi-level feature extraction, 2) a low-resolution space (LR) for learning compact, representative features, 3) a super-resolution space (SR) for upsampling low-resolution features from SR, and 4) a low-high fusion and reconstruction space (LHFR) for final image restoration. Specifically, HR extracts multi-scale high-resolution features and fuses them with low-resolution cues to produce residual images, while LR distills complementary representations from HR to improve restoration quality. To supply LHFR with richer features, SR super-resolves LR outputs before fusion. We further introduce two modules to bridge the high- and low-resolution spaces. The Feature-Refined Correlation Matching Transformation (FR-CMT) module selects the top $C/r~(C~\text{denotes the number of channels;~}r\geq1~\text{controls the squeezing level})$, from the fusion between max- and mean-pooled high-resolution features to replace less informative channels in the low-resolution Transformer. The Adaptive Channel Modulator (ACM) adaptively recalibrates multi-scale high-resolution features, ensuring that only task-relevant information propagates to LR. Extensive experiments demonstrate that UHDformer++ reduces model parameters by at least 86\% compared with recent state-of-the-art methods while achieving substantial performance gains across $5$ UHD restoration tasks, including low-light image enhancement, dehazing, deblurring, deraining, and desnowing. Code will be released at https://github.com/supersupercong/uhdformerplus.

Registering images acquired with different microscopy modalities is essential for relating complementary measurements of the same specimen. In correlative X-ray fluorescence (XRF) and optical microscopy, the XRF map often covers only a small region of an optical image acquired from the same or an adjacent tissue section. Field-of-view (FOV) localization is necessary but can be difficult when appearance and structure differ across modalities. Here we evaluate training-free vision language model (VLM) localization on two datasets representing same-section high-correspondence and adjacent-section low-correspondence imaging. We test unconstrained and metadata-constrained search and compare VLMs with geometric controls, classical template matching, and two alternative training-free approaches (DINOv2 and multiGradICON). Direct VLM prompting produced content-dependent spatial signals but was not reliable alone. Classical matching was most accurate when cross-modal structure was preserved but failed in the low-correspondence collection. A proposal-and-verify workflow used repeated VLM predictions as candidates and image-based similarity to select the final location. This workflow recovered useful localization in the low-correspondence regime.

Current evaluation of computer-use agents is split between long-horizon workflow benchmarks and atomic GUI-grounding tests. This leaves an under-instrumented middle layer: realistic component-centered interactions (e.g., toggle a button set) that are short enough to diagnose and rich enough to capture the burdens of modern interfaces. We present ComponentBench, a benchmark and diagnostic pipeline for component-level evaluation of computer-use agents on modern web UIs. ComponentBench is organized around a library-agnostic ontology of 97 canonical UI components instantiated as 2,910 programmatically verified tasks across widely used component libraries, paired with cleaned human reference trajectories that enable evaluation of both task success and interaction efficiency. Beyond task collection, we introduce a scalable pipeline for auditing realized structural difficulty after implementation and synthesizing structured failure analyses across tasks and component families. Evaluating seven models -- GPT-5.4, Gemini 3 Flash, GPT-5.4 mini, GPT-5 mini, Gemini 3.1 Flash-Lite, Qwen3-VL-235B, and UI-TARS-1.5-7B -- across four observation and action spaces, we show that these design choices critically impact performance. Within a single shared harness, changing only the observation and action space shifts task success by more than 30% for the same model: GPT-5 mini falls from 83.1% with accessibility-tree observations to 48.9% with coordinate-only Pixel control. Moreover, even the fastest configuration takes 3.7x as long as the matched human reference, and spatial manipulations that are trivial for humans continue to challenge current agents.

Single-photon cameras based on single-photon avalanche diode (SPAD) technology are gaining popularity for 3D sensing, thanks to their extreme sensitivity and time resolution. There are two key challenges with single-photon cameras that limit their widespread use: (i) they suffer from non-linear distortions called ''pile-up'' when operated in high-photon-flux conditions, and (ii) they generate a large volume of raw photon data, creating a severe data bottleneck at each sensor pixel. In this work, we show that while compressive capture techniques successfully mitigate data transfer challenges, they exacerbate the effects of dead-time distortion because they fail to retain sufficient information about the photon detection history to allow post-processing pile-up correction via existing methods. We propose a new computational-imaging method that combines free-running capture with an analysis-by-synthesis software pipeline to mitigate pile-up distortions. Our results with hardware emulations and full-scene and single-pixel simulations show that our method can reliably capture scene distance and reflectance over a wide range of illumination conditions. Our work will enable high-resolution SPAD cameras that are severely bandwidth-constrained to operate in real-world high-flux scenarios.

X-ray fluorescence (XRF) microscopy maps elemental distributions, while optical microscopy can provide complementary morphological context. Localizing XRF fields of view (FOVs) in optical images is difficult because the two modalities differ in contrast mechanism and resolution. Most current workflows place each XRF tile independently, even when acquisition metadata already record the tiles' relative scan positions. This study formalizes XRF tile-group localization, in which one optical-frame placement is estimated for the whole group, constrained by acquisition geometry and quantified using group intersection-over-union (GroupIoU). In a controlled case study, independent localization failed with GroupIoU 0.000, whereas group localization achieved 0.931. Replacing the normalized cross-correlation (NCC) metric with mutual information (MI) gave nearly identical results, showing that the outcome is not specific to one local similarity metric. In another multiscale case study, using a coarse XRF survey scan to connect the fine-scale tile group to the optical image increased mean GroupIoU from 0.694 to 0.856. These case studies support using acquisition geometry as an explicit constraint when localizing related XRF tiles.

LLM-as-judge evaluation reduces response quality assessment to a single holistic A/B preference choice, providing no mechanism to isolate which quality dimensions drove the preference or distinguish model errors from genuine label ambiguity. We propose SESSE (Sketch, Expand, Sort, Summarize, Evaluate), a training-free framework that decomposes holistic judgment into structured sub-questions mined directly from the judge's own error cases; requiring no oracle responses, task-specific rubrics, or fine-tuning. On RewardBench (n=1,000), SESSE achieves near-parity with the chain-of-thought baseline and is competitive with RISE-Judge-32B (92.7%), a fine-tuned specialist, while remaining fully training-free. Per-criterion vote evidence provides an interpretable audit trail for diagnosing label ambiguity and judge failure modes unavailable from a single holistic output token.

LLM-as-a-Judge, which leverages a large language model to evaluate natural language generated by another AI application or model, has become a standard, scalable approach for accelerating and extending costly human evaluation. However, most work treats a judge as a static artifact, evaluating it once at construction or against a fixed benchmark. In contrast, we argue that an LLM judge running in a production system is better understood as having a lifecycle: it must be built, trained, deployed, and continuously maintained as the surrounding data evolves, and each phase poses distinct technical and operational challenges. We present such a lifecycle for the LLM judges that evaluate user-facing recommendation explanations at Netflix, where our pipeline generates and the judges assess hundreds of thousands of distinct show-level explanations per week, served across the mobile experience to millions of members. Our framework has four phases: (I) Birth, defining multiple evaluation criteria and building curated benchmark datasets with human labels and rationales; (II) Training, refining the judges' rubrics via Reasoning-Aligned Rubric Tuning (RART), a rubric-tuning procedure that uses a meta-judge over reasoning output as the learning signal; (III) Deployment, in which one judge serves two production roles: quality gating and reflective generation; and (IV) Monitoring, a continuous Human-in-the-Loop alignment process that detects drift and triggers re-tuning behind a human review gate. We report post-launch results from a five-week A/B test over tens of millions of members, in which the judge-aligned explanations shifted member viewing toward novel content (previously unwatched) and increased successful browse-to-play sessions relative to a no-explanation control, with no quality-related takedowns.

As CGM-based AI tools approach clinical deployment, whether their accuracy is equitable across patient demographics remains insufficiently tested. To enable this evaluation, we constructed FairGlucose, a 300-patient CGM cohort balanced across 12 demographic strata (age x gender x type 1/type 2 diabetes), with 132,480 forecasting samples and 3,945 unique behavioral events (meals, exercise, medication) logged by 81 patients. Benchmarking 33 models across four families on 2-hour glucose forecasting, we find that population-level external validation can conceal substantial subgroup disparities. Aggregate out-of-distribution metrics appear stable (approximately 1.0), yet subgroup-level ratios range from 0.8 to 1.4, with T1D patients showing 6 mg/dL higher prediction error than T2D (p < 0.001). This disparity persists across all 33 models, suggesting a property of the prediction task rather than any single architecture. Further analysis shows that subgroup performance gaps align with the proportion of clinically hard cases, and that input-length sensitivity varies across demographics, motivating personalized configurations. Frontier LLMs underperform specialized neural models by 1-6 mg/dL; behavioral events contribute negligibly (approximately 0.1 mg/dL) even under oracle event access. These findings establish that population-level validation alone is insufficient for equity assessment of digital health AI, motivating subgroup-disaggregated reporting as a default standard.